On 14 October, Cohere compared coding-agent teams. A strong lead with 16 cheaper planning workers averaged 61.8% of tests passed, versus 63.2% for an all-strong-model team, at reported costs of $155 versus $510 for the set.

This is not a percentage of completed projects. Fixed-size configurations ran once; the result does not guarantee savings elsewhere.

What the agents are asked to do

The main environment is ProgramBench. Its original paper describes 200 tasks in which an agent reconstructs a program's behavior using the executable original and documentation, without its source code. Hidden tests assess the replacement implementation's behavior. Examples range from command-line utilities to complex systems such as SQLite and FFmpeg.

This changes what planning means. The agent is not merely writing a sequence of steps: it explores how the reference program responds to inputs and edge cases. Discovering the specification is a substantial part of the job. Additional explorers can contribute information rather than simply produce more code.

For contrast, the study also uses Commit0-Lite. The original Commit0 paper describes tasks with an API specification and tests but missing library implementations. This creates a different problem: when requirements are supplied, the benefit of further exploration cannot be assumed to remain the same. These older benchmark papers provide context for the new experiment; they are not presented as this week's releases.

More workers can also mean more integration

Our editorial interpretation treats the finding as a question about the scarce resource. If a team does not understand a program sufficiently, independent behavioral probes may help. If the constraints are already clear but implementation is difficult, dividing exploration into still more pieces may leave the real bottleneck untouched.

There is also a distinction between discovering facts in parallel and changing shared code in parallel. Two observations about different inputs may be straightforward to combine. Two implementations based on incompatible assumptions about formats and errors need reconciliation. This identifies what a local pilot should measure; it does not establish the cause of every difference in the study.

In an illustrative case, a team is rebuilding a small file-conversion utility. Workers could investigate empty files, different encodings and invalid fields, while the lead integrates the findings into one implementation. Success has a concrete meaning: known cases produce the intended results, and unfamiliar cases are assessed with separate tests. Message counts cannot substitute for that evidence.

What the reported cost means

When applying this approach to a real project, we would track computational spending, elapsed time and human review time separately. A cheaper experiment may need more repairs afterwards. A faster run may be valuable despite a larger bill. These are different objectives and should not be compressed into a single efficiency promise.

The useful contribution is a specified organization, task and metric. It supports deliberate allocation of roles rather than a universal instruction to add agents. SIExpo has not reproduced the experiments and reports the figures as Cohere's findings, with the test environment's limitations intact.

Editorial verification: 2026-10-15T16:12:02.996741+00:00 (UTC). Cover: AI illustration, not a documentary photograph.